Guide · made with Zapier
How do you build an AI news monitoring workflow? Nick's six-step competitor monitor
Build a news monitoring workflow in six steps where only one calls a model: it reads the item and says what it means in three sentences.
The short answer
You build an AI news monitoring workflow by letting plain automation do everything with a right answer and calling a model once, to say what the news means. In Nick's competitor monitor, that is six steps with one model call. The job is simple to describe. A competitor ships something, and your team should get a short summary in Slack. The expensive way, which Nick calls the token max version, hands the whole thing to an agent: it reads the feed, decides if the item matters, pulls the record, formats it, writes it up, and posts it. Every one of those is a model call. Nick breaks it apart instead. A new item in the feed is a trigger. Checking it against your competitor list is a string match, and a model can get that wrong. Pulling the competitor's CRM record is a database read. Cleaning dates and fields is formatting. None of those need a token. The one step that does is the read: tell me what this means for our company, in three sentences. Then the result goes to Slack. Build it that way and you pay for judgment only where it happens.
Get the Zapier framework guide →Built with Zapier, the sponsor of this video.
Step by step
Trigger on a new item in the monitoring feed ▶2:04
Start the workflow when a new item lands in the feed you already watch for news. Nick is clear that this is not a judgment call. "That's not a judgment call. That's a trigger." It should cost zero tokens.
String-match against your competitor list ▶2:13
Check whether the item names a competitor you already track. Keep the list of names in one place and compare the item's text against it. If nothing matches, stop the run there. Nick's point is that a model has no advantage on a match like this and can even get it wrong.
Pull the competitor's record from the CRM ▶2:24
Use the matched name to look up that competitor's record in your CRM. This is a database read, so it should be deterministic. The same name should always return the same record.
Clean up the fields before anything reads them ▶2:32
Fix the dates, trim the fields, and handle the general formatting with ordinary automation steps. Nick's view is that none of this is ever worth a token.
Add the one model step: what does this mean for us? ▶2:39
Now hand the cleaned item and the CRM record to a model with one instruction: read this and say what it means for our company, in three sentences. This is the step that needs judgment, and it is the one Nick is happy to pay for. Keep the output short so the summary stays readable in a channel.
Package it and post to Slack ▶2:50
Put the summary together with the item and the competitor's details, and send the whole thing to the Slack channel your team reads. That is delivery, not judgment, so it runs without a model.
Keep the model step swappable ▶3:51
Set up the model step so you can change providers easily. Nick's third rule is to keep model options open, because price and performance shift every couple of months. If swapping is a dropdown and not a project, you can take the win when a better option shows up.
“Okay, now that requires judgment. That part I'm happy to pay for.”
Nick Standlea, in the video
Key moments
- ▶1:38The competitor-monitoring example
- ▶1:46The token max version: one agent does it all
- ▶2:04Step one: a new item is a trigger
- ▶2:13Step two: match the competitor
- ▶2:24Step three: read the CRM record
- ▶2:32Step four: clean the fields
- ▶2:39Step five: the model says what it means
- ▶2:50Step six: post to Slack
- ▶2:54One of six steps needed the model
- ▶4:35Ask your AI which steps could be deterministic
Six steps, one model call. Nick's competitor monitor breaks into six steps. One of them needed the model's judgment. The other five did not.
The token max version calls a model at every step. When one agent reads the feed, decides if the item matters, pulls the record, formats it, writes it up, and posts it, every one of those is a model call. The news monitor is the same job either way. Only the bill changes.
A model can get the easy parts wrong. Matching a competitor's name is a string match. Nick notes that a model has no advantage there and can get it wrong, because it predicts a likely answer instead of running fixed logic.
The paid step is short on purpose. The model step asks for three sentences on what the news means for the company. It is the only place in the workflow where judgment is needed, and the cleanup before it means the model reads tidy fields.
Zapier modeled 71% average savings across 24 workflows. Zapier ran 24 common agent workflows through the same split and found 71% average savings. Its table includes a marketing competitive brief at 90% lower modeled cost. Zapier says these figures are modeled, not taken from live customer workflows.
What does each step of the news monitor cost?
Five of the six steps are rules. The trigger fires when a new item arrives. The match checks the item against a list of names. The CRM step reads one record. The cleanup fixes dates and trims fields. The Slack step posts the result. Each has a right answer you could write down ahead of time, so each can run for zero tokens.
The fifth step is the exception. Deciding what a competitor's news means for your company is judgment, and that is where Nick spends tokens. The result is a monitor that still gives your team an AI-written read on the news, while the model only does the part a model is good at.
Before you build it, mark the steps
Nick's closing advice applies to any monitor you are about to build. Go through it step by step and mark which steps actually need a model. His guess is that it is fewer than you think. Then ask your AI model which steps in the workflow could be deterministic, and compare its answer with your own list.
Method and sources: Claims about the video come from its YouTube transcript, with timestamps taken from transcript start times; the 24-workflow study, the competitive brief figure, and the modeled-only caveat come from Zapier's valuemaxxing page, and the web search detail comes from Zapier's agents guide, checked when this page was drafted.
Sources and references
Related guides
Questions people ask
What is an AI news monitoring workflow?
It watches a feed for news about the companies you care about and sends your team a short read on what it means. In Nick's example, it starts when a competitor ships something, and the output is a summary of what is happening in the market, posted in Slack.
Which step in a news monitoring workflow actually needs AI?
In Nick's build, only one: the step that reads the item and says what it means for the company, in three sentences. The trigger, the competitor match, the CRM lookup, the cleanup, and the Slack post all have right answers and run without a model.
Do I need an AI agent to monitor competitors?
Nick's video argues you do not need one for the whole job. An agent that reads the feed, decides if it matters, pulls the record, formats it, writes it up, and posts it makes a model call at every step. Five of those six steps can be plain automation.
Why not let the model decide whether a news item mentions a competitor?
Because it is a string match. Either the item names a company on your list or it does not. Nick says a model has no advantage there, and in fact it can get it wrong, since a model predicts a likely answer instead of running fixed logic.
What should the prompt in the model step say?
Nick's version is short: read this and tell me what it means for our company, and do it in three sentences. The cleanup steps before it mean the model gets tidy fields and the CRM record, not raw feed text.
Should I add a second agent to check the summary?
Nick warns against the pattern where a team gives an agent work that plain automation could do and then builds a second agent to check the first. That is paying twice for one outcome. Keeping the rule-based steps out of the model in the first place removes most of what would need checking.
How much does building it this way save?
Nick does not give a per-run figure for this workflow. He cites Zapier, which modeled 24 common agent workflows this way and found 71% average savings. Zapier's page also lists an illustrative marketing competitive brief at 90% lower modeled cost, and notes the study is modeled, not based on live customer workflows. Your result depends on how complex the workflow is and which models you use.
Can an AI agent go out and find the news for me?
Zapier's agents guide says its agents can search the web and gather news articles on a topic. Nick's build starts from a feed you already have, so the trigger costs nothing and the model is saved for the reading step.
Which model should I use for the summary step?
Nick does not name one. His rule is to keep model options open and not hardwire a single provider into the stack, because price and performance change every couple of months. Build the step so changing models is a dropdown.
Zapier sponsored this video. Nick chose what to build and what to say.